gh-150942: Optimize PicklerMemoProxy.copy() with _PyDict_SetItem_Take2

[corona10]

• Issue: Improve performance by using reference stealing methods #150942

[corona10]

cc @eendebakpt

1-2% consistenatly faster with simple replace logic, I think that it's worth to do this.

PicklerMemoProxy.copy/2505entries: Mean +- std dev: [base] 136 us +- 4 us -> [opt] 133 us +- 5 us: 1.02x faster
PicklerMemoProxy.copy/50005entries: Mean +- std dev: [base] 4.57 ms +- 0.06 ms -> [opt] 4.52 ms +- 0.05 ms: 1.01x faster

Geometric mean: 1.02x faster

import io
import pickle
import pyperf


def make_proxy(n):
    buf = io.BytesIO()
    pk = pickle.Pickler(buf, protocol=pickle.HIGHEST_PROTOCOL)
    data = [object() for _ in range(n)]
    nested = {i: [object(), object()] for i in range(n // 2)}
    pk.dump(data)
    pk.dump(nested)
    proxy = pk.memo  # PicklerMemoProxy
    # Keep a reference to the source pickler alive via closure.
    return pk, proxy


def main():
    runner = pyperf.Runner()

    for n in (1000, 20000):
        pk, proxy = make_proxy(n)
        entries = len(proxy.copy())
        runner.bench_func(
            f"PicklerMemoProxy.copy/{entries}entries",
            proxy.copy,
        )


if __name__ == "__main__":
    main()

[eendebakpt]

The change itself is ok (the dict is fresh, so no risk of concurrent threads mutating).

But is the proxy used in any computations? I thought the proxy was only for human inspection.

[corona10]

Yeah, I also not sure how it will be practical

[eendebakpt]

Yeah, I also not sure how it will be practical

I am fine with merging though. The code looks cleaner (although this is subjective), and if we do not merge we risk someone else making a similar PR.